Kimi K2.6: The Open-Weight Model That Ties GPT-5.5 on Coding at One-Eighth the Price

TL;DR: Moonshot AI's current flagship is Kimi K2.6 (kimi-k2.6), released April 20, 2026 under a Modified MIT license with open weights on Hugging Face. It's a 1-trillion-parameter Mixture-of-Experts model that activates only 32B parameters per token, ships with a 256K (262,144-token) context, and runs natively in INT4. The benchmarks are genuinely competitive: SWE-Bench Pro 58.6% (a tie with GPT-5.5-class coding), Humanity's Last Exam with tools 54.0%, SWE-Bench Verified 80.2%, and an Artificial Analysis Intelligence Index of 54 — the top open-weights score, alongside MiMo-V2.5-Pro and ahead of DeepSeek V4 Pro. API pricing is $0.60 / $2.50 per million tokens, roughly one-eighth of Claude Opus. Here's what holds up and what to watch.

What Is Kimi K2.6?

Kimi K2.6 is Moonshot AI's flagship as of June 2026, and it replaced the entire earlier K2 line — Moonshot officially discontinued the kimi-k2 series on May 25, 2026, pointing all traffic at K2.6. The model ID is kimi-k2.6 (OpenRouter slug moonshotai/kimi-k2.6).

It is the kind of release that matters more for the open-weights ecosystem than for the absolute frontier. Three things define it:

1. It's open weights you can actually run. The full checkpoint is on Hugging Face under a Modified MIT license — free for commercial use, with the only real string being a visible "Kimi K2.6" credit requirement for products above 100M monthly active users or $20M/month revenue. For everyone smaller, it's effectively MIT.

2. It's a 1T-parameter MoE built for efficiency. One trillion total parameters, but only ~32B activated per token, and it's distributed in native INT4 quantization. That's what makes a trillion-parameter model deployable on hardware budgets that aren't measured in racks.

3. The headline feature is Agent Swarm. K2.6 can fan a task out across up to 300 coordinated sub-agents over roughly 4,000 steps, and Moonshot demos it sustaining 12+ hours of continuous autonomous coding. This is the capability the version bump is really selling.

The lineage is worth a sentence: the open-weight K2 family started with Kimi K2 Thinking (late 2025), then K2.5 in January 2026 added native multimodality and the first Agent Swarm, and K2.6 in April 2026 is the post-training refinement that scaled the swarm and sharpened coding.

What Are the Verified Benchmarks?

A note on sourcing, because it matters more for an open-weights model than for a closed one. Moonshot publishes its own evals, and several are marked in its own tables as re-run under K2.6 conditions rather than cited from third parties. The numbers below come from Moonshot's tech blog cross-referenced against Artificial Analysis, llm-stats, and independent coverage from DeepLearning.AI's The Batch and VentureBeat. Where a figure leans on Moonshot's own harness, I mark it vendor-claimed.

Two rows deserve more than a line.

The AA Intelligence Index of 54 is the one number I'd trust most, because it's Artificial Analysis's neutral harness, not Moonshot's. It puts K2.6 at the top of the open-weights field — tied with MiMo-V2.5-Pro at 54, ahead of DeepSeek V4 Pro at 52 and GLM-5.1 at 51. That's the honest ceiling: best open weights, not best overall. For context, the same index has Claude Opus 4.8 at 61 and GPT-5.5 at 59-60.

SWE-Bench Pro at 58.6% is the load-bearing claim for the "ties GPT-5.5 on coding" framing, and it's the one most worth scrutiny. It's vendor-reported but widely repeated across independent coverage, and it lands right at the GPT-5.5/Opus-4.6-class band on that specific benchmark. Note the gap to the harder closed frontier elsewhere, though: on SWE-Bench Verified, Opus 4.8 posts 88.6% to K2.6's 80.2% — an eight-point spread that's real, even if the price gap is far wider than the capability gap.

One caution on the HLE number. Humanity's Last Exam has had documented score-inflation issues between vendor harnesses and independent re-runs across the industry this year, so treat 54.0% as a tool-augmented, vendor-condition result rather than a settled fact.

How Does It Compare to Claude Opus 4.8, GPT-5.5, and DeepSeek V4?

Where K2.6 sits against the June 2026 frontier, using the figures I could corroborate:

A few honest caveats. K2.6 is not the smartest model on this table — Opus 4.8 and GPT-5.5 both clear it on the aggregate index, and Opus leads SWE-Bench Verified by eight points. What K2.6 wins is the intelligence-per-dollar-with-open-weights corner. Among models you can download and run yourself, it leads the field. Against DeepSeek V4 Pro — the other serious open-weights contender — K2.6 edges ahead on the index (54 vs 52), but DeepSeek's V4 Flash tier is dramatically cheaper at $0.14/$0.28, so for high-volume, error-tolerant work DeepSeek still owns the bottom of the cost curve.

The positioning that holds across sources: Opus 4.8 is the intelligence leader, GPT-5.5 is the all-rounder, and Kimi K2.6 is the open-weights coding-and-agent leader you can self-host. For a fuller cross-model breakdown see the FrankX models tracker, the Claude Opus 4.8 analysis, and the DeepSeek V4 analysis.

What's the Pricing?

The pricing is the whole pitch. At $0.60 input / $2.50 output per million on Moonshot's official API, K2.6 is roughly one-eighth the cost of Claude Opus 4.8 on input and one-tenth on output. On a per-output-token basis it's about 42x cheaper than Opus. That's the lever: if your evals show K2.6 clears your quality bar on coding or agentic tasks, the cost delta is large enough to change what you can afford to build.

And the API price is only the floor, because the weights are free. Self-hosting trades API cost for infrastructure and ops cost — a trade that pays off only at volume, but one that's available at all precisely because this is an open-weight release.

The Open-Weight and Self-Host Angle

This is where K2.6 is genuinely differentiated from Opus 4.8 and GPT-5.5, neither of which you can run on your own hardware at any price.

The full checkpoint is on Hugging Face under a Modified MIT license. The modification is narrow: products above 100M monthly active users or $20M/month in revenue must show a visible "Kimi K2.6" credit. Below those thresholds — which is to say, almost everyone — it behaves like a permissive MIT license, including for commercial use and fine-tuning.

The architecture is built to make self-hosting plausible rather than theoretical:

• 1T total parameters, ~32B active per token. You pay compute for the active path, not the full trillion, on each forward pass.
• Native INT4 quantization. The weights ship pre-quantized, which cuts the memory footprint substantially versus a naive FP16 trillion-parameter model.
• 256K context (262,144 tokens) for input, with output capacity up to the same 262,144-token window per the API specs.

The practical reasons to self-host are the usual ones, now actually attainable at near-frontier quality: data residency and privacy (nothing leaves your VPC), no per-token metering on high-volume workloads, the ability to fine-tune on proprietary data, and no vendor lock-in. The honest counterweight: running a 1T MoE well is a real infrastructure project, and for most teams the $0.60/$2.50 hosted API will be the right call until volume or compliance forces the move.

What Is Agent Swarm and What Is K2.6 Best At?

Agent Swarm is K2.6's signature capability and the clearest reason to reach for it over a general-purpose model. It lets the model decompose a task and fan it out across up to 300 coordinated sub-agents over roughly 4,000 steps, running in parallel. K2.5's first-generation swarm capped at 100 sub-agents and ~1,500 tool calls — K2.6 roughly triples the sub-agent ceiling and nearly triples the step budget.

Crucially, Moonshot is explicit that the architecture didn't change between K2.5 and K2.6. The swarm gains are post-training, not architectural — more compute spent on long-horizon stability, instruction following, and the routing that coordinates the swarm. The model got better at using a capability it already had. That's a useful thing to know, because it means the headline number (300 sub-agents) is a behavioral ceiling, not a hardware one, and your mileage will depend heavily on how well-specified your task is.

The one number that cleanly isolates the swarm: BrowseComp jumps from 78.4% (K2.5) to 86.3% (K2.6) in swarm mode — a +7.9-point gain on agentic web research, which is exactly the kind of fan-out-and-reconcile work the swarm is designed for. The Terminal-Bench 2.0 jump from 50.8% to 66.7% tells the same story on the coding side.

What K2.6 is best at:

• Long-horizon agentic coding — the 12+ hour autonomous-coding demos and the Terminal-Bench gain point here.
• Parallel web research and synthesis — Agent Swarm + BrowseComp 86.3% is the strongest swarm-specific evidence.
• Coding-driven UI/UX generation — Moonshot positions it for prompt-and-visual-to-interface work across Python, Rust, and Go.
• Cost-sensitive, high-volume agentic pipelines where you want near-frontier coding quality without frontier pricing.

What it's not: the absolute reasoning frontier. If your work is dominated by the hardest single-shot reasoning where a silent error is expensive, Opus 4.8 or GPT-5.5 still earn their premium.

What Does It Mean for Builders?

For developers

The migration story is simple: if you're already routing to an OpenAI-compatible endpoint, K2.6 is kimi-k2.6 on Moonshot's API or moonshotai/kimi-k2.6 on OpenRouter, and the request surface is conventional. The thing to actually do is run your own coding evals before believing the 58.6% SWE-Bench Pro tie — vendor-reported benchmarks are a starting hypothesis, not a deployment decision. Sweep K2.6 against your current model on your real tickets and let the cost delta argue its case.

For agentic and long-horizon work

This is where K2.6 earns its keep. Agent Swarm is the differentiator, but it rewards a well-specified first turn — give it the full task definition, clear success criteria, and let the swarm fan out, rather than feeding it piecemeal. The 12-hour autonomous-coding claim is real in Moonshot's demos but assumes a gradeable definition of "done"; build that rubric before you turn it loose.

For cost-conscious routing

K2.6 changes the routing math for the middle tier. For coding and agentic work that previously justified an Opus- or GPT-5.5-tier call, K2.6 at $0.60/$2.50 may now clear your quality bar at a fraction of the cost — and if it doesn't quite, DeepSeek V4 Flash at $0.14/$0.28 sits below it for the genuinely error-tolerant, high-volume cases. The discipline is the same as always: match the model to the task's cost-of-error, not to the leaderboard. Reserve the closed frontier for the routes where a silent miss is expensive.

For teams with compliance or data-residency constraints

This is K2.6's unique unlock. It's the first time a near-open-weights-frontier coding-and-agent model is available to run entirely inside your own environment under a permissive license. If "the data cannot leave our VPC" has been blocking you from frontier-class coding assistance, K2.6 is the most credible answer on the board today.

FAQ

Is Kimi K2.6 better than Claude Opus 4.8?

No, not on raw intelligence. Opus 4.8 leads the Artificial Analysis Intelligence Index (61 vs 54) and SWE-Bench Verified (88.6% vs 80.2%). What K2.6 wins is value and openness: it's roughly one-eighth the price, you can self-host it under a permissive license, and on SWE-Bench Pro it lands in the same band as much pricier models. For the absolute frontier, Opus; for open-weight coding-and-agent work at a fraction of the cost, K2.6.

How much does Kimi K2.6 cost?

$0.60 per million input tokens and $2.50 per million output tokens on Moonshot's official API — roughly one-eighth the input cost and one-tenth the output cost of Claude Opus 4.8. Via OpenRouter it's $0.684 / $3.42 with the routing markup. The open weights are free to download and self-host under a Modified MIT license.

Can I self-host Kimi K2.6?

Yes. The full checkpoint is on Hugging Face under a Modified MIT license, free for commercial use (the only condition is a visible "Kimi K2.6" credit for products above 100M MAU or $20M/month revenue). It's a 1T-parameter MoE with ~32B active per token, distributed in native INT4, which makes self-hosting practical — though running a trillion-parameter MoE well is still a real infrastructure project.

What is Agent Swarm and how does it differ from K2.5?

Agent Swarm lets K2.6 decompose a task across up to 300 coordinated sub-agents over roughly 4,000 steps, running in parallel — up from K2.5's 100 sub-agents and ~1,500 tool calls. The gain is post-training, not architectural: the model got better at coordinating a capability it already had. The cleanest evidence is BrowseComp rising from 78.4% to 86.3% in swarm mode.

What's the context window and which benchmarks are verified vs vendor-claimed?

The context window is 256K (262,144 tokens), with output capacity up to the same window. The most trustworthy number is the Artificial Analysis Intelligence Index of 54 (neutral third-party harness, top of the open-weights field). The coding and reasoning figures — SWE-Bench Pro 58.6%, HLE-with-tools 54.0%, AIME 96.4% — are Moonshot's own evals, widely cited but not all independently reproduced. Treat the AA index as settled and the rest as vendor-claimed until third parties confirm.

Kimi K2.6 vs DeepSeek V4 — which open-weights model should I use?

K2.6 edges DeepSeek V4 Pro on the Artificial Analysis Intelligence Index (54 vs 52) and is the open-weights coding-and-agent leader. DeepSeek's V4 Flash tier is far cheaper ($0.14/$0.28), so for high-volume, error-tolerant work DeepSeek owns the bottom of the cost curve. Pick K2.6 for agentic coding and swarm workloads; pick DeepSeek Flash for cheap high-volume throughput. See the DeepSeek V4 analysis for the full picture.

Analysis by Frank — former Oracle AI architect who helped build Oracle's AI Center of Excellence, now building agentic systems independently and making music with AI. Published June 5, 2026 with benchmarks validated against Moonshot AI's tech blog, Artificial Analysis, llm-stats, and independent coverage. Vendor-claimed figures are marked as such.